AI MODELS
Follow important model changes, understand what changed, and check the official evidence.
New versions, capabilities, access and real-world deployments.
UnifoLM-WLA
Unitree released the Base checkpoint and fine-tuning code.
Previously reported
September 20: model modules and action-expert training code available
This announcement
September 28: Base weights and fine-tuning code available
UnifoLM-WLA
Unitree adds model modules and action-expert training code.
Previously reported
September 11: UnifoLM-ER-1 and ER-Flow weights released
This announcement
September 20: model modules and action-expert training code released
Figure Helix
Figure reports tests across 30 unseen homes.
Previously reported
Helix 02: data collected in deployment environments
This announcement
Helix 2.5: three whole-body tasks in 30 unseen homes
Search models, developers or robots; explore versions, updates and official sources.
| Model series | Research organization / team | Type | Resources | Recorded versions | Version release date | Related robots | Actions |
|---|---|---|---|---|---|---|---|
| UnifoLM-WLA Unitree Robotics’ humanoid foundation model family for tabletop and whole-body manipulation. | Research organization / team | TypeVLA | Resources | Recorded versionsUnifoLM-WLA-1.0-Base | Version release dateSeptember 28, 2026 | Related robotsG1· Evaluated on | Actions |
| Figure Helix Figure AI vision-language-action system for humanoid robots. | Research organization / team | TypeVLA | Resources— | Recorded versionsHelix 2.5 | Version release dateSeptember 17, 2026 | Related robotsFigure 03 | Actions Choose a version to compare |
| Gemini Robotics Vision-language-action models for robot motor control. | Research organization / team | TypeVLA | Resources— | Recorded versionsGemini Robotics 2 | Version release dateJuly 30, 2026 | Related robotsALOHA 2· Official compatibilityApollo (2023)· Official compatibility | Actions Choose a version to compare |
| Gemini Robotics On-Device VLA models optimized to run locally on robotic devices. | Research organization / team | TypeVLA | Resources— | Recorded versionsGemini Robotics On-Device 2 | Version release dateJuly 30, 2026 | Related robotsALOHA 2· Official demonstrationApollo (2023)· Official demonstration | Actions Choose a version to compare |
| MolmoAct2 Ai2 action-reasoning model family for real-world robot deployment. | Research organization / team University of Washington 2 institutions | TypeVLA | Resources | Recorded versionsMolmoAct2MolmoAct2-Think | Version release dateMay 5, 2026May 5, 2026 | Related robotsSO-100SO-101 | Actions Choose a version to compare |
| WALL-OSS WALL-OSS is a vision-language-action model family from X Square Robot. The official research page dates it to 2025-09-08; its paper was submitted to arXiv on 2025-09-15. | Research organization / team | TypeVLA | Resources | Recorded versionsWall-OSS-0.5WALL-OSS Flow | Version release dateMay 28, 2026— | Related robots— | Actions Choose a version to compare |
| GR00T NVIDIA foundation model family for cross-embodiment robot control. | Research organization / team | TypeRobot foundation modelVLA | ResourcesWeights
| Recorded versionsGR00T N1.7 | Version release dateApril 17, 2026 | Related robotsG1· Official compatibility· Official demonstration· Evaluated onAI Worker+1 | Actions Choose a version to compare |
| π₀ Physical Intelligence vision-language-action model family. | Research organization / team | TypeVLA | Resources | Recorded versionsπ₀.₇ | Version release dateApril 16, 2026 | Related robotsG1· Community port | Actions Choose a version to compare |
| UnifoLM-VLA Unitree’s Vision-Language-Action model family for humanoid manipulation. | Research organization / team | TypeVLA | Resources | Recorded versionsUnifoLM-VLA-0 | Version release dateJanuary 29, 2026 | Related robotsG1· Evaluated on | Actions |
| LingBot-VLA Robbyant VLA pretrained on 20,000 hours from nine dual-arm configurations. | Research organization / team | TypeVLA | Resources | Recorded versionsLingBot-VLA | Version release dateJanuary 27, 2026 | Related robots— | Actions |
| EVO1 Evo-1 is a lightweight vision-language-action model from the MINT lab at the School of AI, Shanghai Jiao Tong University. It preserves semantic alignment through two-stage training. The paper was submitted on 6 November 2025, with Tao Lin as first author. | Research organization / team | TypeVLA | Resources | Recorded versionsEVO1 | Version release dateNovember 6, 2025 | Related robots— | Actions |
| X-VLA Soft-prompted cross-embodiment VLA family. | Research organization / team2 institutions | TypeVLA | Resources | Recorded versionsX-VLA 0.9B | Version release dateOctober 12, 2025 | Related robotsWidowX-250 6DOF | Actions |
| Genie Operator Genie Operator-1 is the AgiBot World team’s vision-language-action model, pretrained on robot manipulation data using latent action representations. Weights, fine-tuning code and inference examples are available, alongside GO-1 Air without the latent planner. | Research organization / team | TypeVLA | ResourcesWeights | Recorded versionsGenie Operator-1GO-1 Air | Version release dateSeptember 19, 2025— | Related robots— | Actions Choose a version to compare |
| RDT Robotics Diffusion Transformer family for bimanual manipulation. | Research organization / team | TypeVLARobot policy | Resources | Recorded versionsRDT2-FMRDT2-VQ | Version release dateSeptember 2025September 2025 | Related robots— | Actions Choose a version to compare |
| EO-1 EO-1 is a unified embodied foundation model for general robot control, combining interleaved vision, text and action training with multimodal reasoning. The paper lists Shanghai AI Laboratory, Fudan University, AgiBot and Northwestern Polytechnical University. | Research organization / team | TypeVLA | Resources | Recorded versionsEO-1 | Version release dateAugust 28, 2025 | Related robotsWidowX-250 6DOFG1 | Actions |
| SmolVLA Compact VLA family for accessible robot learning. | Research organization / team | TypeVLA | Resources | Recorded versionsSmolVLA base | Version release dateJune 2, 2025 | Related robotsSO-100SO-101 | Actions |
| OpenVLA Open vision-language-action model family for robot manipulation. | Research organization / team | TypeVLA | Resources | Recorded versionsOpenVLA 7B | Version release dateJune 13, 2024 | Related robotsWidowX-250 6DOF | Actions |
| Octo Open-source generalist robot policy family. | Research organization / team | TypeVLA | Resources | Recorded versionsOcto-Base 1.5Octo-Small 1.5 | Version release dateMay 20, 2024May 20, 2024 | Related robotsWidowX-250 6DOF | Actions Choose a version to compare |
| RT-X RT-X is a cross-embodiment robot policy family from the Open X-Embodiment Collaboration, led by Google DeepMind. RT-1-X is a non-VLM policy with public code and weights; RT-2-X uses a VLM, with 55B and 5B research configurations. | Research organization / team | TypeRobot policyVLA | Resources | Recorded versionsRT-1-XRT-2-X | Version release dateOctober 2023October 2023 | Related robots— | Actions Choose a version to compare |
| InternVLA InternVLA is a Shanghai AI Laboratory model series spanning robot manipulation and navigation. M1 studies spatially guided policies, A1 and A1.5 combine understanding and action, and N1 uses a dual-system navigation design. | Research organization / team | TypeVLA | ResourcesWeights
| Recorded versionsInternVLA-A1InternVLA-A1.5InternVLA-M1InternVLA-N1 | Version release date———— | Related robots— | Actions Choose a version to compare |
— means no confirmed information; it does not mean unsupported or unavailable.
